Abstract
Medical imaging AI systems typically quantify uncertainty for isolated image-level predictions, yet clinical use depends on how these uncertain predictions affect downstream diagnostic and management decisions. We introduce a sequential conformal safety framework that propagates image-derived uncertainty through multi-stage, guideline-based clinical decision pathways and provides distribution-free, finite-sample safety guarantees. Modeling patient management as a staged decision tree, we prove that, with probability at least , at least one clinically safe pathway remains reachable under uncertainty. The framework composes stage-wise conformal validity with explicit miscoverage allocation to enable calibrated pruning without eliminating all safe alternatives. We instantiate the method in ophthalmology, modeling diabetic retinopathy management with macular edema as a latent state resolved through OCT imaging, and validate on glaucoma workflows. Experiments on five public datasets demonstrate calibrated pathway-level retention and low operational decision risk while maintaining interpretable, non-vacuous decision sets. These results extend conformal prediction from isolated outputs to structured clinical workflows with formal safety guarantees.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/UNSURE2026_076.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=~Alejandro_Sanchez_Guinea1
BibTex
@InProceedings{GuiAle_Sequential_MICCAISAT2026,
author = { Guinea, Alejandro Sanchez AND Erokhina, Mariya},
title = { { Sequential Conformal Safety for Trustworthy Clinical Decision Pathways in Ophthalmology } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17260},
month = {pending},
page = {pending}
}